Add user-suppliable batch_evaluator argument to minimize/maximize

[hmgaudecker]

What

Threads a batch_evaluator keyword from the public minimize/maximize API through
create_optimization_problem and the OptimizationProblem container to
InternalOptimizationProblem, replacing the hardcoded process_batch_evaluator("joblib")
in _optimize. Users can now pass either a built-in name ("joblib", "pathos",
"threading") or a callable conforming to the BatchEvaluator protocol.

Also:

• Removes the dead batch_evaluator Literal field from the Tranquilo / TranquiloLS
  dataclasses. It was never read — the evaluator lives on the optimization problem, not
  the algorithm — so setting it silently no-op'd.
• Exports BatchEvaluator from the top-level namespace so users can type their custom
  evaluators against the protocol.

Why

The batched criterion evaluations a trust-region optimizer like tranquilo requests were
locked to the in-process joblib evaluator. Some workloads need a different execution
backend — e.g. a distributed, MPI-backed pool that fans each batch of candidate points
out across nodes/GPUs for an expensive simulation-based criterion. There was already a
process_batch_evaluator resolver and a BatchEvaluator protocol; this just lets the
caller reach them.

This is also the foundation for a follow-up: a first-class executor-backed
(concurrent.futures / mpi4py.futures) distributed evaluator, where minimize itself
parks non-root ranks in a worker pool so the optimizer runs on a single driver. That
addresses a real footgun — under MPI, srun starts N copies of the program, so a naive
"run minimize on every rank with a collective evaluator" design ends up with N
optimizers kept in lockstep, which diverges under floating-point nondeterminism. This PR
is the minimal, backward-compatible plumbing that makes the safe design expressible.

Tests

Adds test_minimize_uses_custom_batch_evaluator,
test_maximize_uses_custom_batch_evaluator (a spy callable is the one optimagic actually
invokes), and test_minimize_rejects_unknown_batch_evaluator (clear error on an unknown
name).

Backward compatibility

Fully backward compatible — batch_evaluator defaults to "joblib", the prior behavior.

---

Happy to add a CHANGES.md entry referencing this PR number under the appropriate
next-version heading (0.5.4 vs 0.6.0 — your call).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
src/optimagic/__init__.py	93.75% <100.00%> (+0.20%)	⬆️
...imagic/optimization/create_optimization_problem.py	88.28% <100.00%> (+0.09%)	⬆️
src/optimagic/optimization/optimize.py	92.10% <100.00%> (ø)
src/optimagic/optimizers/tranquilo.py	99.13% <ø> (-0.02%)	⬇️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.